Matching Inconsistently Spelled Names in Automatic Speech Recognizer Output for Information Retrieval

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 451–458, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Matching Inconsistently Spelled Names in Automatic Speech Recognizer Output for Information Retrieval Hema Raghavan and James Allan Department of Computer Science University of Massachusetts Amherst Amherst, MA 01003, USA {hema,allan}@cs.umass.edu Abstract Many proper names are spelled inconsis- tently in speech recognizer output, posing a problem for applications where locating mentions of named entities is critical. We model the distortion in the spelling of a name due to the speech recognizer as the effect of a noisy channel. The models fol- low the framework of the IBM translation models. The model is trained using a par- allel text of closed caption and automatic speech recognition output. We also test a string edit distance based method. The ef- fectiveness of these models is evaluated on a name query retrieval task. Our methods result in a 60% improvement in F1. We also demonstrate why the problem has not been critical in TREC and TDT tasks. 1 Introduction Proper names are key to our understanding of topics in news. For example, to determine that a news story is on the 2004 elections in the United States, the words President Bush, John Kerry and USA are necessary features of the story. In other words, names of people, places and organizations are key entities of a news story. For many tasks, like in topic detection and tracking (TDT), the entities form an important feature for distinguishing topics from one another. For example, it is the people that distinguish stories on the 2004 election from stories on the 2000 U.S election. Names, especially rare and foreign ones are a problem for automatic speech recognition (ASR) systems as they are often out of vocabulary (OOV) i.e., they do not exist in the lexicon of the ASR system. An OOV word is replaced with the most similar word in the lexicon of the speech recognizer. Sometimes, even if a name is in the lexicon of the speech recognizer, it may have multiple spelling variants. The following is a sample ASR snippet from the TDT3 1 corpus that demonstrates how the same entity may have different spellings even within the same snippet of ASR text. ...newspaper quotes qaddafi is saying they’ll turn them over but only if they’re allowed ..leader moammar gadhafi says he doesn’t want an interna- tional confrontation over the suspects in the.. In this work, we aim to find methods by which to cluster or group names in ASR text. We evaluate a variety of techniques that range from a simple string-edit distance model to generative models using both intrinsic and extrinsic evaluations. We get statistically significant improvements in results for ad-hoc retrieval when the query is just the name of a person. We also explain why the problem of misspelled proper names in ASR has not been an issue in the TREC spoken document retrieval (SDR) track or in topic …

Content truncated to 500 words